Columns

But the fact remains that even if a GPU is programmed to do a particular task relatively inefficiently, it only has to be less than ten times as inefficient and it can still blow the socks off anything Intel or AMD has to offer. And while the GPUTeraSort team used raw OpenGL to program their GPU, some Microsoft research, published in the same month as the GPUTeraSort paper, has come up with a programming API called Accelerator, which does much of the hard work of translating standard-style code into GPU fodder. Benchmarks written with Accelerator include 2D image-processing algorithms, which in some cases exceed ten times the speed of a 3.2GHz Pentium 4.

It's possible this will remain a niche area for research and scientific computing, but that would be a shame. With all the research going on, and with the number of PCs sporting dual 16x graphics slots, the time may come when you'll have the option of popping in a second graphics card to take the burden off your main CPU.